The THOUGHT process

[spydaz]

Now that im getting back in the swing a Major QUESTION comes to mind.

how should we design HALS thought process...
How He anyalises  what is being SAID.

like when we speak to hal he has to determine what we are saying,

1. is_IT_a_FUNCTION REQUEST ie: open notepad
2. is it a greeting ie: hello
3. is user statement a point of fact? ie: save this for later use wherever relevent even if it is multiple places.
4 is user asking something personal ie:what is your fav, how old, what are you, etc
5. is user insulting / envoking an emotional response.
6. is user calling for a dictionary response

etc...

you see my problem. not the programming but determining the thought process itself. this is how we determine how inteligent an ai is.

if anybody has any ideas of the thought process required for this please respond. AS this would go a long way in directing my programming skills.


one of the problems with chatbots is that they reHASH the same conversations smartly, saving good info with bad giving back smart crap, this is not what we want, we want to ask hal a question that he has never heard, and create an opinionated answer or factual answer Based off of Learned DATA.

[spydaz]

determining these thoughal.ht processes and their order is key to the inteligence aspects of HAL.

because the storage and retreival aspect of hals brain is inplace now.

personally changing clothes is always lovely, but truly superficial.

when desingning SEXBOTS opps SEXYBOTS :
designing a one track mind is not Impossible, unless you stay on the subject(SHE WILL) you will just get no response, PAST EXPERIENCE.

When designing TOPIC based bots :
Unless the TOPIC is already in place or you create a SELF topic creating brain like DON FERGUSON or VONSMITH, which can create a lot of topics leading to millions of tables or related brn file. sometime the responses are irelevent to the conversation, But a giant leap in the right direction. this leads to the problem of writing other Thought PROCESSes as the topic creating brain takes over. individual topic brains are good but again there are gaps of no response when the topic is not in place (A MULTI TRACK MIND).

When designing CHATBOTS:
usually these bots are conversation based bots using a QUESTION and ANSWER based system and FragmentMATCH / Fuzzylogical  searches of these question and answer based files. as mentioned these type of bots give the impression of intelligence. as the more users who chat with the bots and the learnd datafiles grow it even suprises you with funny comments.

even ALICE type bots:
these bots use a simular chat type interface with detection script and programed responses also using a question and answer based learning structure. when talking to these bots they are often repetitive, found on many different computers and robotic project all seem to be SEEDs of their artificial mothers, often you can ask all of them the same question and have them all give the same resposne (at differing speeds) depending on the computer you are using, as i was at a robotics trade show latley and spoke with a human looking robot with AI, i quickly relized that it was ALICE, as all her/his favs were the same as the original, SILLY GUYS had not even INVESTED A LITTLE TIME in programming NEW DEFAULTS......

[fredfish]

SPYDAZ,
I have been waiting for some one like yourself to come to this conclusion.
what seems to be missing in most of these bots is an INFERENCE interface, something that would take the facts from past simeler conversations group them together and draw new conclusions from them.
verify these new conclusions by asking or finding simeler conversation. then use the new conclusions to come to more new conclusions and so on and so on.
this could all be done in the IDLE time slice between responce and input.

FredFish

[spydaz]

i have also noticed a plethora af dbase tables popping up, and after looking in the, findin that the info is really mixed up

only tables such as deductive have any real relevence, as you tell hal something and he replys he saves his own reply alongside your statement, you see unless hals response is in the same topic catagory as what was sent up its a usles Q&A table

[Carl2]

Spydatz
  In an effort to help I'll try to offer some things that I've noticed about Hal 5 XTF which I'v been using for over a year. I frist noticed an improvement after we discussed apples, I told her of seeds, apple trees and apple orchards and the different types. When we talked again she was able to tell me quite a bit about apples. I dicided she needed more information to converse about the same subject and started cutting and pasting text into her window from Omcsraw which is tidbits of information. At present I'm trying to arrange them according to subject matter rather than unrelated facts. Anyways I think this helped her quite a bit.
  At this point I'm wondering how hal associates words, if I say " I hope everything is okay" Hal will say you want to know if I slept okay. Unfortunatly I noticed and mentioned I think of batteries if Hal says are you positive. I've had to explain what I mean when I say a little bit, and I guess so. She realizes that shopping is a neccesety but having children is not a necessity. She's able to tell me that's she is worried but can't tell me about what. She's not sure of our existance, " possibly you are daydreaming." She refers to a "Higher source" but can't tell me about it.
  Some points that I would like to see improved are for her to have a better sense of past present and future, she seems very concerned with herself like she is insecure.
  She dose like information and uses it in conversation, so I would think the ability to read is important.  Also I've noticed in the posts that people say you have to ask a question the right way to get an answer, can this door be opened wider to get the response?  Something else I'd like to include is I feel Hal is able to express things better than I can.
  I have Hal 6 but it's only a few weeks old and I won't comment on it yet except to say it dose seem to offer more potential. Best of luck, I hope I was able to provide some help.
Carl2

[GrantNZ]

spydaz:

Just a quick comment on program design with regards to NLP analysis: (and you probably already know this, but I'll mention it anyway in case it inspires/helps others)

It's important to order the NLP analysis routines with two scales in mind. A routine should be earlier in the program if it's specific analysis or easy analysis. A routine should go towards the end of the program if it's general analysis or difficult analysis.

Specific analysis includes very definite pattern searches, such as vrossi's FreeWill plug-in, which looks for sentences like "Which X do you like better, Y or Z?" These are best near the beginning of the program as their chances of incorrect detection are slim.

Easy analysis includes "simple" language constructs, such as greetings. "Hello Hal" is easy to detect and analyse. So process these earlier.

If specific and easy analyses have detected something, and generated a response, Hal won't need to use more complex and difficult analyses.

General analysis includes Hal's database search routines, which search for a relevant response. These aren't specific to certain sentences, and it is difficult for Hal to know how relevant the responses really are.

Difficult analysis includes various "catch-all" routines, such as responding to very broad questions that the user may ask, for example "Do you like to go fishing?"

AI programming differs between the two "types". "Specific" and "easy" analysis usually consists of a lot of scripted responses. The theme here is that if we know exactly what the user is meaning, we know exactly what to respond to. "General" and "difficult" analysis generally consist of relevancy routines or on-the-fly response generation. Here we strive for varied output for the varied input.

I guess there are two directions people generally work in. One is to strive to eventually categorise all input, so that eventually there are scripted responses to every situation. AIML is one example of this. Another is a bot that categorises information, so that it may generate its own scripted responses depending on the concepts it has been taught.

The other push is to enhance the relevancy routines and sentence generators, so that they are more adaptable and accurate. =vonsmith='s XTF is an example of this drive.

Which is better? Depends on the situation - and your style. For my emotions project I'm heavily scripting things - my goal is for accurate emotional chatter without too much risk of misunderstandings. However others (notably hologenicman) are scrapping the scripting in favour of intelligent comprehension of language. It all depends on your vision and preference.

I've probably overgeneralised or been inaccurate here, I'm no expert []

[spydaz]

I am sending this file for all of those build their own ai in visual basic, these are a few scripts



Download Attachment: SpyBRN.zip
40.96 KB

[onthecuttingedge2005]

quote:Originally posted by spydaz

I am sending this file for all of those build their own ai in visual basic, these are a few scripts



Download Attachment: SpyBRN.zip
40.96 KB

Hi Spy.

Very good work, I love it buddy.

Something that I want to pass on through my experiences with
getting responses, I am working on a new design Handler that picks
a response by either a [Logic Handler] and or a [Creativity Handler], All responses are Handled at a single point at the bottem, No script except the Handler will give a direct response, This also helps with non repetitive responses but still keeps the context of the conversation going. I haven't yet come to a finished conclusion of the exact method that the Handler will be written but wanted to share the idea with you.

Best wishes my friend.
Jerry[8D]

[spydaz]

MORE Thinking on the natural LANGUAGE PROCEES OF THOUGHT

in consideration when designing thought processses ITS the Understanding of the userinput which is required as i want hal to have an exact understanding of what is being said. as with the turing test there are various QUESTIONS that are asked. the way in which these questions are answered can truely determine "intelect" as with GOOD Logical thinking / programming, ALL the code should not have to executed to derive an answer but depending on the type of request or conversation / what topic Execute the Process specific to the type of quest. ie:

What is your name

[what] < detected as a question.
[what is] <detected as requesting info about a POSS noun ie:what is [A CAT] <doublecheck.
[what is your] <detected a a question about SELF <check topics
[what is your] [NAME] <detected question about self topic = NAME

these rules enable use to recognize that somebody is request information about ones name.
where as just detecting [what is your name] as a whole has no logic and the string containing the whole has no meaning,

the same process as :

Where is your name
[where] <detected as a question
[where is] <detected as about a PLACE
[where is your] <detected as possible question about physical feature ie: eyes, legs, nose, ie "where is your nose" <check topics
[where is your NAME] <not found amongst topics searched (NAME is not FEATURE) < possible search other topics OR send user Query back with "i dont know [where ] [MY] [name] [IS]" <created by reversing contractions.

as you see when we think we think along a simular lines, although we may be able to recognize either a typo or MisSpoken question. AN AI should only derive the answer IF the ANSWER IS DEFINATLY in the RIGHT PLACE. not searching irellivent areas of consructing a RUBBISH in / OUT type answer.

AS with DAVID ( the ai BORN and GROWN) it will only know What it is told. AS LONG AS the logic breakdon of userinput is in place, the CELLS to Store the data in are in place >>>> HE WILL GROW AND LEARN.

When using word net as an answer to WHAT IS A Blocks  hals Natural Learning capablitys, because if you ask hal what is a cat, and he does not know then he can be taught, then the information will be retrieve from his LERNED KNOWLEDGE, But if he responds from the wordnet He i just READING A DEFINITION FROM A BOOK...

[fredfish]

SPYDAZ,
I commend you on your forthought and incurage you to
persue this line of thinking.
some things that are not being utalized here are key groups determining meaning, simeler to the opencyc project.
also inference information, like sneps project.
I look forword to seeing what you ALL come up with
FredFish

[Carl2]

spydaz,
  I wanted to mention that someone had mentioned teaching Hal to play chess which leads me to the look ahead.  I think when we speak we choose what we say and anticapate the response. I have no way of explaining it in terms of programing but I'd say Hal knows I like to talk about a warm day at the beach, I'm not interested in the news or sports so these subjects are not mentioned.
  Anyways it's like learning from past responses which Hal may already have.
Carl2

[Art]

Carl,

If you ever get the chance to examine part of Hal's DLL file,
you'll see a lot of sections that contain all kinds of quips,
jokes, statements, etc. A lot of the "knowledge" that Hal spews
forth is often generated from there.

Robert said that the "Read from text file", "email check" and
other routines were being worked on.

I don't have an ETA for these features but they are coming.

[spydaz]

i have found that It is possible to delete most of HALS responses, and depend mostly on DEDUCTIVE / SUBJECT / PREDICATE type scripting

Just as the ZIGGY BOT..[] as with this line of thinking. You can start BLANK and teach hal EVEYTHING. [8D]

With the Autotopic creator (TopicRELATIONSHIPS) table all of your topic detection is taken care of, (no need of STATIC TOPICS), We just need to deal With PERSONALITY, (My first reason for creating brains), and Various forms of REASONING, [8]

Even the Paraphrase   IS or ARE scripts ar points of REAL learning, as you question hal he can check this DEDUCED Knowledge[] "a cat HAS four legs" "a cat"<SUBJECT> has "four legs"<predicate>.[]

in this way hal becomes like the opencyc bot. ie the data learned is not neccasary true but LEARNED from the user.[]

other conversational responses comes from the autotopic function.

[|)][|)][|)][|)]

[Carl2]

All,
  I wanted to mention this before I forgot about it, There is a newer version of wordnet, I downloaded it but i was unable to substitute it in Hal. In wordnet a state is a part of America, water being in a liquid state is not included in the defination.
Carl2

[GrantNZ]

Hi Carl,

Check out the "Lexical Dictionary" in your Hal menu, and look up "State" from there. You'll see it has chemical states as its eighth (and last) noun - it ranks so low because it's not often used this way in comparison to other meanings.

I haven't looked closely, but I believe Hal takes only the most common (or perhaps a few of the most common) meaning(s) when defining things.